HABAEK: High-Performance Water Segmentation Through Dataset Expansion

다국어 초록 (Multilingual Abstract)

Water segmentation is a critical component of effective disaster response and water resource management. Authorities increasingly employ high-resolution, near-field photography from drones, CCTV, and smartphones to monitor rivers, lakes, and reservoirs, enabling proactive management in agriculture and conservation. However, these images present unique challenges due to variable visual characteristics such as reflections, lighting, and complex backgrounds. Deep models like CNNs, U-Nets, and transformers process large data but demand high computation, limiting real-time use. This research addresses these challenges by demonstrating an efficient, data-centric approach to maximize the performance of an existing model rather than designing a new complex architecture. We focus on SegFormer, a Transformer-based model that often results in a weaker inductive bias compared to CNNs, making it heavily dependent on extensive training data to learn fundamental spatial patterns. To compensate for this, we propose Habaek, a model that enhances SegFormer’s generalization capabilities by strategically augmenting its training with diverse water-centric datasets, including ADE20K and RIWA. To maintain efficiency, we adopt a head fine-tuning strategy, which significantly reduces computational load while preserving high performance. The Habaek model achieves state-of-the-art results, with IoU improving from 0.9198 to 0.9439, and shows superior F1, recall, accuracy, and precision compared to existing models. This study ultimately proves that for modern, low-bias architectures, a data-centric strategy combined with efficient fine-tuning is a highly effective and practical path to achieving benchmark performance in critical segmentation tasks.